Research On Perception And Grasping Technology Of Intelligent Robot Based On Deep Learning

Three-dimensional perception technology has a wide range of applications in industrial production and daily life.At present,industrial robots are mainly used in specific or relatively stable working environments to repeatedly execute the taught actions to complete the established work tasks.In recent years,artificial intelligence technologies such as computer vision and deep learning have been developed rapidly,which provide technical support for industrial robots to realize three-dimensional perception and intelligent motion.This thesis use binocular vision technology and deep learning technology to perform 3D reconstruction and intelligent perception of target objects and combine them with industrial robot grasping applications,and mainly accomplish the following work:Firstly,the theoretical models of monocular and binocular cameras and camera aberration models are studied,and an experimental platform for parallel optical axis binocular vision is designed and built.A three-step optimization method of threshold restriction,reliable parallax filling and bilateral filter processing is proposed for the shortcomings of the SGBM semi-global stereo matching algorithm,and the point cloud data is reconstructed in 3D using binocular vision,and the point cloud data is preprocessed and localized to expand the data set required for training for deep learning environment perception.Secondly,the basic principles and representative networks of deep learning perception model are analyzed,the feature extraction process of deep learning perception model is described,and a suitable deep learning network is selected to carry out the study of intelligent robot environment perception.Then,GA-PointNet++,a deep learning network architecture based on graph attention mechanism,is designed for 3D reconstructed point cloud data of environmental objects.By introducing the self-attention coefficient and neighborhood attention coefficient mechanisms,different neighborhood point cloud data are given different attention weights.The results of perceptual accuracy and robustness experiments show that the designed network architecture achieves 94.7%and 91.3%recognition accuracy on different datasets test sets,respectively;the proposed algorithm can maintain good robustness for point cloud data with fewer number of points,and the proposed algorithm can accurately perceive the target object point cloud data with good network performance.Finally,an experimental platform of intelligent robot perception and grasping system consisting of ABB robotic arm,binocular vision system and deep learning perception system was designed.The Eye-to-Hand model is used to complete the hand-eye calibration,and the Euler angle and quaternion conversion relationship between the robot arm and the target object is established,and the target object position information is sent to the robot.The 3D reconstructed point cloud data is input into the pre-trained model for point cloud recognition,while the ABB robot arm perceives and grasps the experiments,and the experimental results show that the method used in this thesis can achieve accurate and stable recognition and grasping of the target object.